silkylove / Pytorch-ImageSegmentation

Pytorch-ImageSegmentation

A ImageSegmentation Pytorch Implementation of various models and datasets

Introduction

Still working in progress! Trying to include more models and datasets in a similar and simple way.

Datasets

To train and validate the network, this repo use the augmented PASCAL VOC 2012 dataset which contains 10582 images for training and 1449 images for validation.

Training

Make sure to look at the config.py and change it as you like, then run main.py Finally you will get the ckpt and log in checkpoints/

Run 
tensorboard --logdir log 
to see the training log.

TODO

• Unet, PSPNet, DeepLabV3+
• Resnet backbones
• VOC2012 dataset
• More models including Unet-like Resnet based AutoEncoder ...
• More backbones including SE-ResneXt101, Xception and Mobile net
• Deformable CNN
• More datasets, eg, COCO

Results untill now

Now I had done the experiments on deeplabv3+ and pspnet (one can get the training log under ckpt): Performance with 4 1080ti: Deeplabv3+ on VOC2012 validation with 16 batchsize and 30K iterations. The pretrained model can be got from deeplabv3+

Backbone	train/val OS	mIoU paper	mIoU repo
Resnet101	16	78.85%	78.21%

For MobileNet_V2 with Deeplabv3+:

Backbone	Dataset	train/val OS	mIoU paper	mIoU repo	Speed
MobileNet_V2	VOC2012	16	--	69.39%	10ms(512x512)
MobileNet_V2	CitySpcases	16	--	65.27%	15ms(512x1024)

PSPNet on VOC2012 validation with 16 batchsize and 30K iterations in 10 hours. The pretrained model can be got from pspnet

Backbone	mIoU paper	mIoU repo
Resnet101	--	77.61%

Resnet101 based Unet (imgsize 256) on VOC2012 validation with 16 batchsize and 30K iterations in 10 hours. The pretrained model can be got from unet_ae

Backbone	mIoU repo
Resnet101	74.77%

Requirements

Python(3.6) and Pytorch(1.0.0) is necessary before running the scripts. To install the required python packages(expect PyTorch), run

pip install -r requirements.txt